NVIDIA DGX A100

NVIDIA DGX A100. It is a purpose-built artificial intelligence supercomputer and data center appliance introduced by NVIDIA in May 2020. The system is designed to consolidate the workloads of training, inference, and analytics into a unified AI infrastructure, powered by the company's new Ampere architecture. As a flagship product in the NVIDIA DGX series, it represents a significant leap in computational performance and efficiency for enterprise and research AI applications.

Overview

The system was unveiled by NVIDIA CEO Jensen Huang during a virtual keynote at the GPU Technology Conference in 2020. It serves as the foundational building block for larger-scale deployments like the NVIDIA DGX SuperPOD, which aggregates multiple units into a clustered supercomputer. The design philosophy centers on delivering unprecedented AI compute density, aiming to replace entire racks of legacy servers with a single, integrated appliance. This consolidation is intended to simplify data center operations for organizations engaged in advanced research and development across various scientific and commercial fields.

Hardware specifications

At the core of the appliance are eight NVIDIA A100 Tensor Core GPUs, fabricated using TSMC's 7 nm process. These processors are interconnected via NVIDIA NVLink and NVSwitch technology, providing a high-bandwidth, unified memory space. The system incorporates six NVIDIA ConnectX-6 HDR InfiniBand/Ethernet adapters for high-speed networking, crucial for scalability in multi-system clusters. It is built on a server motherboard designed by NVIDIA and includes AMD EPYC CPUs for host processing. Storage is provided by multiple NVMe solid-state drives, and it is housed in a 4U rack unit chassis with a specialized liquid cooling system to manage thermal output.

Software and ecosystem

The system ships with the pre-installed NVIDIA DGX Software Stack, which includes the NVIDIA Base Command platform for multi-user, multi-system management and job scheduling. This software layer provides optimized container images via the NVIDIA NGC catalog, featuring frameworks like PyTorch, TensorFlow, and MXNet. It also includes the NVIDIA Magnum IO suite for accelerated I/O and the CUDA toolkit for parallel computing. Full integration with Kubernetes and support for Red Hat Enterprise Linux ensures compatibility with modern data center orchestration and enterprise IT infrastructure.

Applications and use cases

Primary deployments are found in large-scale scientific research, such as at the Argonne National Laboratory for COVID-19 research and the University of Florida for an academic AI initiative. It is extensively used for natural language processing models like GPT-3, computer vision tasks, recommendation systems, and genomics research at institutions like the King Abdullah University of Science and Technology. The system also enables complex simulations for autonomous vehicle development at companies like Toyota and accelerates drug discovery pipelines in the pharmaceutical industry.

Market context and competition

Upon its launch, the system positioned NVIDIA against other high-performance computing vendors like AMD with its Instinct GPUs and Intel with its Xeon CPUs and Habana Labs accelerators. It also faced competition from integrated cloud computing AI services offered by Amazon Web Services, Microsoft Azure, and Google Cloud Platform. The product's success reinforced NVIDIA's dominance in the AI hardware market, influencing the development of subsequent architectures like Hopper and shaping procurement strategies for national projects like the EuroHPC JU and the United States Department of Energy's exascale computing efforts.

Category:NVIDIA hardware Category:Artificial intelligence Category:Supercomputers Category:2020 introductions